Skip to content


Track: Simulation

Harnessing Deep Reinforcement Learning to Coordinate Automated Guided Vehicles

Monday, April 12, 1:45-2:25pm EDT

We present a deep reinforcement learning (DRL) policy that was trained to control a fleet of Automated Guided Vehicles (AGVs). The policy determines the loading and drop off tasks of AGVs as they move products through a factory. The goal of the policy is to maximize product throughput. Notably, the policy shows a 50% improvement in throughput over a shortest-queue heuristic, while lowering the AGV utilization by 15%. We attribute the improvement to the policy’s intelligent management of congestion. Interestingly, the policy chooses to “hide” AGVs away from the center of the factory floor, resulting in lower AGV utilization and lower congestion. Our presentation will detail the adaptation of the AGV simulation as a DRL problem, the policy training process and the quantitative results. Additionally, we will discuss how the same workflow can be readily applied to a variety of other use cases.

Johnny Davenport image

Johnny Davenport

Johnny Davenport

Reinforcement Learning Scientist at Pathmind

Johnny is a reinforcement learning scientist at Pathmind who has solved dozens of industrial, supply chain and green energy logistics problems with reinforcement learning. He has a Ph.D. in physics from the University of California: Santa Cruz, where he studied electronic phenomena in atomically thin crystals. Johnny enjoys applying the mathematics of condensed matter physics to deep learning problems.